betaFeedback

WorksComputer Science

NN-Based FeFET Compact Modeling for SPICE Simulation

Abstract

Prototyping and early-stage circuit evaluation for emerging transistor technologies require fast and accurate compact models for SPICE simulations. Standard physics-based compact models, such as BSIM, offer high fidelity but demand extensive domain expertise and time-consuming equation formulation for every new physical effect, significantly slowing technology development. Neural network (NN)-based models offer a scalable alternative approach, but they rely heavily on smooth numerical solver data with explicit high-order derivative information. However, measurement data suffers from complex noise patterns, making derivative extraction challenging and complicating NN-based transistor modeling. This work presents a derivative-free NN training framework employing spectral normalization to naturally smooth derivative behavior and filter measurement noise. Using drain current measurements from a fabricated 28 nm ferroelectric device, we design an NN-based compact model that captures the full spectrum of polarization amplitudes and supports dynamic polarization switching in transient simulations. To demonstrate its capability, we deploy the model in a commercial SPICE software to simulate write and read operations for large-scale memory arrays with up to 512x512 size, including parasitic effects. Our approach bridges raw measurements with NN-based SPICE-compatible transistor modeling, significantly reducing model development time and enabling rapid circuit-level prototyping for emerging devices.

Introduction

Circuit simulations in the rapidly evolving semiconductor industry demand accurate and fast transistor models. Berkeley short-channel IGFET (BSIM) [1] covers a wide range of modern transistor technologies and remains the most commonly used industry-standard model for analog and digital circuit designs. BSIM relies on physics-based equations to accurately capture underlying transistor behavior, enabling the model to achieve excellent performance in modeling device characteristics. However, developing an analytical solution to accurately describe the underlying physics of device operation is often time-consuming and requires substantial domain expertise, which significantly slows down the development process for emerging technologies since all new physical effects need to be properly studied, characterized with equations, and implemented in the model.

Neural network (NN)-based transistor modeling received significant attention over the recent years as a promising alternative to conventional models. NNs can effectively create a standalone, compact model for specific working conditions dictated by the range of data and input parameters used for NN training. Matrix multiplication, prevalent in many NN architectures, is a simple operation that enables NN-based models to significantly reduce model complexity and accelerate SPICE simulations.

While NN-based transistor models have demonstrated success across various devices [2, 3, 4, 5], prior implementations largely relied on synthetic data from numerical solvers (e.g., SPICE tools) that provide high-precision derivative information. However, experimental measurement data is inherently noisy, and in addition to standard Gaussian noise, there are also various anisotropic, conditionally dependent, or structured noise patterns, making the extraction of derivative information with conventional numerical and filtering methods challenging. Since derivative smoothness and stability are essential for SPICE solver convergence, fitting continuous, noise-resilient NN models to measured data remains a significant hurdle. Therefore, a reliable way to handle measurement data with NN-based transistor models has yet to be found.

Our main contributions within this paper are as follows:

(1) We propose a derivative-free approach for NN-based transistor model training using spectral normalization, which guides NN derivative behavior and filters noise present in the measurement data.

(2) We measure transfer characteristics of a 28 nm Ferroelectric FET (FeFET) devices and develop an accurate NN-based transistor model, which covers the entire range of polarization amplitudes and supports dynamic polarization changes in transient simulations.

(3) For comprehensive testing, we deploy the NN-based FeFET model within a commercial SPICE tool and simulate the write and read operations across various memory array configurations. The size of arrays varies from 2×22 \times2 up to 512×512512 \times512, and all arrays incorporate parasitic elements.

Background

Ferroelectric-Based Transistors

Adding a thin (e.g., 10 nm) HfO2HfO_2 Ferroelectric (FE) layer to a MOSFET gate stack yields a CMOS-compatible FeFET [6, 7] (see Figure 1(a)). The electric polarization of the FE layer interacts with the channel below, repelling or attracting carriers depending on the polarization direction. Since polarization of the FE layer is permanent, it modulates channel carriers to

FeFET gate stack and IDS-VGS sketch

Figure 1. (a) FeFET gate stack with the added HfO2_2-based FE layer. (b) Sketch of the IDSI_{\mathrm{DS}}-VGSV_{\mathrm{GS}} characteristics of FeFET with intermediate states. At VGS=0V_{\mathrm{GS}} = 0 V the current is OFF, even in the low-VTHV_{\mathrm{TH}} state, while the ON current at VRV_{\mathrm{R}} (e.g., 1 V) depends on the polarization state.

This work was supported by the German Research Foundation (DFG) Device-Circuit Co-Optimization for Reliable FeFET-based Brain-inspired Computing in the Face of Variability (DeCOP) under Grant AM 534/8-1 and Grant 547377347. The project on which this report is based was sponsored by the Federal Ministry of Research, Technology and Space under the funding code SWC3.0 16IS23069 “RAVEN”. Responsibility for the content of this publication lies with the authors. (Corresponding author: Hussam Amrouch)

Rodion Novkin, Simon Thomann, and Hussam Amrouch are with the Technical University of Munich; TUM School of Computation, Information and Technology; Chair of AI Processor Design; Munich Institute of Robotics and Machine Intelligence, 80333 Munich, Germany (e-mail: {rodion.novkin,s.thomann,amrouch}@tum.de). form two distinct threshold voltage (VTHV_{\mathrm{TH}}) states for non-volatile memory (NVM) storage (see Figure 1b). Negative polarization maximizes VTHV_{\mathrm{TH}} (high-VTHV_{\mathrm{TH}} state associated with “0”), while positive polarization minimizes it (low-VTHV_{\mathrm{TH}} state associated with “1”). Gate pulses exceeding the coercive field (e.g., ±4\pm4 V) switch the polarization state by creating the necessary electric field across the FE layer, which is stronger than the coercive field. Additionally, modulating the pulse amplitude or width achieves partial polarization, shifting VTHV_{\mathrm{TH}} to intermediate states for multi-level cell (MLC) operation [8].

FeFET-Based Memory Arrays

Using FeFET, dense random access memory arrays can be implemented. Following a design from the literature, we show the architecture of the memory array in Figure 2a at the example of a 3×33 \times3 array [9], where FeFETs form a grid-like structure and serve as memory cells.

FeFET memory array architecture, write scheme, and read scheme

Figure 2. (a) FeFET memory array architecture at the example of a 3×3 array. (b) FeFET memory array write scheme, targeting the top right FeFET. (c) FeFET array read-out bias scheme for the last row.

Figure 2b shows the write scheme to change the polarization of a specific FeFET. It shows all biases per line to write the top-right FeFET (target), while the states of the other FeFETs are not disturbed. The targeted FeFET perceives the full potential difference of the write voltage VWV_{\mathrm{W}} across the FE layer, with its source and drain terminals grounded. To write the value to the target FeFET while not disturbing the state of other FeFETs in that row, a 1/31/3-based disturbance-free write scheme is employed [9]. For this, bitline (BL) and signal line (SL) of the non-target columns are biased with 2/3⋅VW2/3 \cdot V_{\mathrm{W}} to reduce the strength of the electric field below the coercive field. Similarly, the wordline (WL) of the non-target rows needs to be biased with 1/3⋅VW1/3 \cdot V_{\mathrm{W}} to not disturb their states.

To read out a stored data word as shown in Figure 2c, a row is selected by applying VRV_{\mathrm{R}} to a WL, which enables it. By applying VRV_{\mathrm{R}} also to the BLs, the current per BL depends on the state of enabled FeFET and encodes the stored value. Using an analog-digital-converter (ADC), the BL currents can be converted back into digital values for further processing.

Device Measurement

To collect the measurement data that forms the foundation of the training dataset, we have measured a 7 μm×10 μm7\,\mu\mathrm{m} \times10\,\mu\mathrm{m} (width ×\times length) FeFET fabricated in a 28 nm process (see Figure 3). Using an MPI TS3000-SE wafer probe station, we have loaded a 300 mm wafer onto the thermally controlled chuck, which is held at 300 K. For electrical characterization, we use the Keysight B1500A semiconductor device analyzer mainframe. To drive the biases and measure currents, the 4 installed Keysight B1511B source measurement units (SMUs) are employed and connected via triaxial cables to the micro positioners, which contact the pads on the wafer [10, 11].

Figure 3. Measurement setup for the electrical characterization of the FeFET on 300 mm wafer.

To gather data on multiple polarization states, a two-step measurement protocol is used. 1) The FeFET is polarized to the target state. For this, we reset the polarization to the high-VTHV_{\mathrm{TH}} state with a 2 s long −4-4 V pulse. Starting from this known state, we polarize the target intermediate state by applying the positive voltage pulse (VWV_{\mathrm{W}}) of different amplitudes for 2 s, which results in achieving a corresponding voltage polarization amplitude state (VP=VWV_{\mathrm{P}} = V_{\mathrm{W}}). The polarization step concludes by waiting for 10 s to allow accumulated traps at the channel-FE interface to detrap. 2) After the FeFET has been polarized, the IDSI_{\mathrm{DS}}-VGSV_{\mathrm{GS}} characteristics are measured. For this, we sweep VGSV_{\mathrm{GS}} from −0.2-0.2 V to 1.051.05 V in 10 mV steps at a fixed VDSV_{\mathrm{DS}} bias and grounded VSV_{\mathrm{S}}. To have more data close to VDS=0V_{\mathrm{DS}} = 0 V, we use a non-linear step size for the VDSV_{\mathrm{DS}} biases. For 1 mV up to 10 mV the resolution is 1 mV, from 10 mV to 100 mV the resolution is 10 mV, and from 0.1 V to 1 V the resolution is 25 mV.

The used amplitudes during the polarization step are 0 V, 1.5 V, 1.8 V, 2 V, 2.3 V, 2.5 V, 2.7 V, 3 V, 3.5 V, 4 V and 4.5 V, i.e., 11 IDSI_{\mathrm{DS}}-VGSV_{\mathrm{GS}} datasets in total. Due to the polarization scheme described above, the 0 V case represents the high-VTHV_{\mathrm{TH}}, since no positive-voltage pulse followed the reset.

Our Derivative-Free NN Training Framework

NN Architecture and Training Algorithm

Although NN-assisted FeFET modeling was explored before [12, 13], a SPICE-compatible NN-based compact model was not introduced. A conventional approach for NN-based transistor model training demands using first- and, sometimes, second-order derivative information and incorporating it into the loss function [2, 3, 4, 5]. However, in contrast to previous works, our training data was measured and not generated by numerical simulators. Therefore, the ground truth for clean and smooth derivatives does not exist. It is possible, in theory, to filter the data, but data filtering in three dimensions (gate and drain voltages, polarization amplitude) is notoriously hard because values in all three dimensions are interconnected and are supposed to represent the same smooth and continuous function IDS=f(VDS,VGS,VP)I_{\mathrm{DS}} = f(V_{\mathrm{DS}}, V_{\mathrm{GS}}, V_{\mathrm{P}}). Since measurements were performed as IDSI_{\mathrm{DS}}-VGSV_{\mathrm{GS}} sweeps by setting VDSV_{\mathrm{DS}} and polarization amplitude VPV_{\mathrm{P}} constant, a measurement drift occurs in these dimensions, leading to noisy IDSI_{\mathrm{DS}}-VDSV_{\mathrm{DS}} and IDSI_{\mathrm{DS}}-VPV_{\mathrm{P}} sweeps. Having axis-aligned or sweep-dependent noise characteristics poses additional challenges, since most filtering methods are designed for isotropic, independent, and identically distributed Gaussian noise, which is not suitable for our data. Therefore, we will explore a derivative-free and filtering-free approach for NN-based transistor modeling.

Since we only have measurement data for positive VDSV_{\mathrm{DS}}, we assume a symmetric device and force this symmetry with the drain current conversion function as:

IDS=VDS⋅eyI;yI=NN(∣VDS∣,VGS,VP)(1)I_{\mathrm{DS}} = V_{\mathrm{DS}} \cdot e^{y_I}; \quad y_I = \mathrm{NN}(\lvert V_{\mathrm{DS}}\rvert, V_{\mathrm{GS}}, V_{\mathrm{P}}) \tag*{(1)}

This way, the sign of the current will be determined by the sign of the VDSV_{\mathrm{DS}}, while the value of the current will be determined from its absolute value. We keep all voltage values and polarization amplitude values unchanged and do not normalize them. Additionally, we omitted data points with IDSI_{\mathrm{DS}} below 1×10−12 A1 \times10^{-12}\,\mathrm{A}, as they are approaching the noise floor of the measurement devices and introduce significant uncertainty.

Having no derivatives available during training leaves us with the following form of the loss function:

LI=a⋅Er(IDS)+Er(yI)(2)L_I = a \cdot Er(I_{\mathrm{DS}}) + Er(y_I) \tag*{(2)}

where ErEr is a L1 loss. Essentially, NN is only trained with the margin loss to capture the logarithmic behavior of the curve, and a part that scales the drain current in a saturation region with coefficient aa. The training starts with a=0a=0 and ends with a=3×106a=3 \times10^6 to ensure a sufficient contribution.

While the loss function in Equation (2) is relatively simple due to the removal of derivatives, it lacks all terms that enable the correct derivative behavior. Our baseline NN architecture consists of three hidden linear layers, each with 16 neurons, and a tanh activation function to limit the layer outputs. This architecture guarantees C∞C^{\infty}-continuity by itself, but it still does not prevent NN from giving wavy and extremely unstable derivatives even for a good IDSI_{\mathrm{DS}} fit. Therefore, an efficient and reliable way to guide NN derivatives has to be introduced.

Spectral Normalization

Originally proposed by Miyato et al. [14], spectral normalization is widely used in the discriminator of Generative Adversarial Networks (GANs), enabling stable adversarial training and mitigating model instability without the need for extensive hyperparameter tuning or implementation of gradient penalties. Spectral normalization is a weight normalization technique designed to control the Lipschitz constant of NN layers by bounding their matrix spectral norms. It reparameterizes the network weight matrix WW by dividing it by its largest singular value σ\sigma (spectral norm):

WSN=Wσ(W)(3)W_{\mathrm{SN}} = \frac{W}{\sigma(W)} \tag*{(3)}

This transformation prevents NNs from producing disproportionately large output swings in response to out-of-distribution inputs. Therefore, spectral normalization is often used for GAN stabilization, adversarial robustness, and NN architecture stability. In this work, we employ scaled spectral normalization to mitigate measurement noise and prevent chaotic gradient behavior in the learned drain current data, since the desired NN output cannot be physically achieved in the case of a regular spectral normalization. Therefore, we multiply the output of a normalized linear layer by a coefficient λ\lambda, which amounts to simple weight scaling. Even though the coefficients λi\lambda^i for each layer ii can be made learnable, it is not recommended, as optimizers often take the path of least resistance, leading to high values of λi\lambda^i. Instead, coefficients are balanced manually with the minimum possible values to cover the entire range of the output. For the employed NN with three hidden layers, λ1,2=1.0\lambda^{1,2}=1.0 (base spectral normalization), λ3=2.0\lambda^3=2.0, λOUT=4.0\lambda^{OUT}=4.0, while the input layer is not normalized.

Since scaled spectral normalization targets NN weights, for inference purposes, the weights of the pre-trained NN can be simply recalculated (fused with a scaled normalization layer), resulting in no additional runtime cost for the model.

Linking Ferroelectric Polarization and Gate Voltage

A physical FeFET has only four terminals: drain, gate, source, and body. A polarization amplitude state VPV_{\mathrm{P}}, used as NN input, is an internal state of the FeFET which can be achieved by applying voltage pulses to transistor terminals. Therefore, a way to link polarization dynamics with standard transistor terminals has to be developed. In this paper, we employ a simplified model of the polarization behavior and assume that the change in the polarization state is driven by the potential difference VGDV_{\mathrm{GD}} between the transistor gate and drain terminals. Additionally, we implement certain logic and boundaries that occur in real physical FeFET devices:

(1) Positive polarization occurs only when VGDV_{\mathrm{GD}} exceeds a certain polarization voltage threshold VPTH. If VGDV_{\mathrm{GD}} is below the threshold VPTH, FeFET operates as a usual transistor and produces the current based on its polarization state.

(2) FeFETs cannot be polarized above the level which the applied potential difference VGDV_{\mathrm{GD}} allows.

(3) Already polarized FeFETs cannot be polarized further unless the applied potential difference VGDV_{\mathrm{GD}} is higher than the one this FeFET was polarized with previously.

(4) FeFET cannot be polarized indefinitely and can only achieve a certain polarization amplitude VPMAX regardless of the applied potential difference VGDV_{\mathrm{GD}} and its duration.

(5) Negative polarization occurs when the potential difference VGDV_{\mathrm{GD}} is negative and its absolute value exceeds VPTH. Polarization amplitude cannot go below the high-VTHV_{\mathrm{TH}} state.

(6) Polarization rate depends on the value of the applied voltage difference VGDV_{\mathrm{GD}}.

Statements from 1 to 5 are boundaries and conditions that need to operate in the time domain. Therefore, they are implemented using various gating functions combined into a differential equation. To derive a polarization speed VPS, we employ the formula for the FeFET switching time [15]:

VPS=dtS=dt0exp⁡[−αkBT⋅1(VGD−V0)2](4)\mathrm{VPS} = \frac{d}{t_{\mathrm{S}}} = \frac{d}{t_0}\exp\left[-\frac{\alpha}{k_{\mathrm{B}}T}\cdot\frac{1}{(V_{\mathrm{GD}}-V_0)^2}\right] \tag*{(4)}

where kBk_{\mathrm{B}} is Boltzmann’s constant, TT is the temperature, t0t_0, α\alpha, and V0V_0 are the minimum switching time, exponential constant, and voltage offset, respectively. dd is the selected unit distance, which can be a unit of polarization amplitude or a normalized polarization unit. For simplicity, we assume that positive and negative polarization speeds have the same dependency with respect to the absolute value of VGDV_{\mathrm{GD}}. Since there is a denominator in Equation (4) which can be equal to zero when VGD=V0V_{\mathrm{GD}}=V_0, this formula cannot be used in Verilog-A directly. Additionally, the function in Equation (4) is unbounded and may be prone to overflows during calculation. Therefore, we propose the following form for the speed function, which captures the dependency between polarization speed and VGDV_{\mathrm{GD}}:

VPS=VPRA1⋅exp⁡[A2⋅tanh⁡(−A3⋅∣VGD∣+A4)−A5](5)\mathrm{VPS} = \frac{\mathrm{VPR}}{\mathrm{A1}\cdot\exp\left[\mathrm{A2}\cdot\tanh\left(-\mathrm{A3}\cdot\lvert V_{\mathrm{GD}}\rvert+\mathrm{A4}\right)-\mathrm{A5}\right]} \tag*{(5)}

The proposed function captures all necessary properties of the original function in Equation (4): a non-linear exponential dependency of the VPS with respect to VGDV_{\mathrm{GD}}, asymptotically limited output when VGDV_{\mathrm{GD}} rises to infinity, or approaches V0V_0 and goes to 1/infinity1/\mathrm{infinity}. Infinite values cannot be simulated, but they can be limited to values small enough not to influence the simulation output. A1-A5 are function fitting coefficients responsible for the dependency between polarization speed and VGDV_{\mathrm{GD}}, and VPR is a polarization rate coefficient that can scale polarization speed. All polarization-related parameters can be found in Table I.

AbbreviationDescriptionDefault
VPMODEnables (>0>0) or disables polarization calculation1.0
VPTHVoltage threshold to initiate polarization1.5
VPMAXMaximum achievable polarization amplitude4.5
A1-A5Coefficients for polarization speed functionFigure 4
VPRPolarization rate coefficient1.0
VP0Initial polarization level (normalized)-1.0
VPDELTASmoothing coefficient for gate/ceiling functions0.001

Table I. Polarization-related parameters in the NN-based FeFET model.

To demonstrate how the proposed in Equation (5) VPS function fits into data from the real device, we employ the timing data for a different 28 nm FeFET [15]. We take the switching speed for ERS transitions and recalculate it into polarization speed using unit distance d=1d=1. Figure 4 demonstrates the resulting fit, which follows the data pattern while limiting the output values of the VPS function. Although there are some differences with the fit from the original function, the primary objective of the VPS function proposed in Equation (5) is to establish a simulation-compatible relationship between VGDV_{\mathrm{GD}} and polarization speed VPS.

Halid-plot of polarization speed function

Figure 4. Halid-plot of the proposed in Equation (5) polarization speed function for ERS data points of the 28 nm FeFET [15] in comparison to the original in Equation (4). Fitting parameters: VPR=1, A1=0.15, A2=10, A3=0.8, A4=1.4, A5=3.5.

Evaluation of nn-based FeFET Model

Since NN was trained on noisy measurement data, any standard quantitative metric such as mean absolute error (MAE) or mean absolute percentage error (MAPE) will not yield meaningful results, as the final value will be heavily skewed by noise. Therefore, in addition to a regular visual evaluation, we also employ a statistical method to estimate the NN fit quality. Figure 5 illustrates IDSI_{\mathrm{DS}}-VGSV_{\mathrm{GS}} (top) and IDSI_{\mathrm{DS}}-VDSV_{\mathrm{DS}} (bottom) fits for VP=4.5 VV_{\mathrm{P}}=4.5\,\mathrm{V}. Minor deviations between NN fit and measured data are the result of the corrections that NN has to perform with three-dimensional data. Measurement data consists of independent IDSI_{\mathrm{DS}}-VGSV_{\mathrm{GS}} sweeps, which are coherent in one dimension only. However, NN has to fit the data as a single smooth and continuous function IDS=f(VDS,VGS,VP)I_{\mathrm{DS}}=f(V_{\mathrm{DS}},V_{\mathrm{GS}},V_{\mathrm{P}}) where all three dimensions are interconnected. Therefore, IDSI_{\mathrm{DS}} values for an entire dimension or even a 2D plane may have to be shifted to achieve the required 3D fit. One of the prime examples for this behavior is the IDSI_{\mathrm{DS}}-VPV_{\mathrm{P}} fit shown in Figure 6. During training, only 11 points shown in the figure for VPV_{\mathrm{P}} were used. Each of these 11 points is represented by a function IDSVP=f(VDS,VGS)I_{\mathrm{DS}}^{V_{\mathrm{P}}} = f(V_{\mathrm{DS}}, V_{\mathrm{GS}}). Thus, if measurements for one of VPV_{\mathrm{P}} points (see VP=2.7V_{\mathrm{P}} = 2.7 V or 4.04.0 V) are noticeably off, the entire IDSVPI_{\mathrm{DS}}^{V_{\mathrm{P}}} plane has to be corrected for this particular VPV_{\mathrm{P}}, which inevitably leads to small mismatches between the NN fit and measurements. A key advantage of forcing the NN to learn a unified, continuous function is its effect on derivative quality. As illustrated in Figure 5, both the transconductance gmg_{\mathrm{m}} and output conductance gDSg_{\mathrm{DS}} are free from measurement noise artifacts and exhibit smooth and stable behavior.

IDS-VGS and IDS-VDS fits

Figure 5. IDSI_{\mathrm{DS}}-VGSV_{\mathrm{GS}} (a-c) and IDSI_{\mathrm{DS}}-VDSV_{\mathrm{DS}} (d-f) fits for VP=4.5V_{\mathrm{P}} = 4.5 V demonstrate a good visual match with the measurement data while having smooth and stable derivatives. Lines are NN fit, circles are measurement data. Measurement data on logarithmic scale plots is not shown since it matches visually with the fit.

IDS-VP fits

Figure 6. IDSI_{\mathrm{DS}}-VPV_{\mathrm{P}} fit for VDS=VGS=0.9V_{\mathrm{DS}} = V_{\mathrm{GS}} = 0.9 V demonstrates a noticeable measurement drift for some polarization amplitudes (e.g. VP=2.7V_{\mathrm{P}} = 2.7 V and VP=4.0V_{\mathrm{P}} = 4.0 V) that was corrected by NN. Line is the NN fit, circles – measurement data.

In addition to a comprehensive visual evaluation, which includes all IDSI_{\mathrm{DS}}-VGSV_{\mathrm{GS}} and IDSI_{\mathrm{DS}}-VDSV_{\mathrm{DS}} sweeps for all polarization states, we also employ a statistical method to estimate the NN fit correctness. A reliable way to estimate the fit is to evaluate the log-residuals between the measurement data and the NN fit in all data points:

R=log⁡(IDSmeas)−log⁡(IDSpred)(6)R = \log(I_{\mathrm{DS}}^{\mathrm{meas}}) - \log(I_{\mathrm{DS}}^{\mathrm{pred}}) \tag*{(6)}

Despite the anisotropic structure of the raw measurement noise, Figure 7 demonstrates that the resulting log-residuals approximate a zero-mean normal distribution. Thus, the residual error between the NN predictions and the measurement data behaves as zero-mean Gaussian noise. This confirms that the NN captures the underlying deterministic behavior for the drain current while filtering out measurement noise. By constraining the network’s Lipschitz constant via spectral normalization, the model avoids overfitting to axis-aligned artifacts, effectively reducing complex measurement noise to uncorrelated Gaussian residual errors, achieving a smooth and unbiased fit.

Log-residual histogram

Figure 7. Log-residuals between measured data and NN fit resemble a normal distribution, which indicates that NN fit filtered out noise present in measurement data and isolated the transistor’s true physical behavior.

SPICE Simulations with nn-based FeFET Model

Although there are custom frameworks for performing SPICE-based FeFET array simulations [16], all of our simulations were performed with a regular commercial SPICE simulator using our NN-based Verilog-A FeFET model. We explore various FeFET memory arrays with different sizes and include parasitics represented by small resistance and conductance in each word, bit, and source lines for all FeFETs. Therefore, an M×MM \times M memory array netlist has M2M^2 transistors and 6⋅M26 \cdot M^2 parasitic elements. The values for parasitic resistance and capacitance were set to 0.5 Ohm and 0.5 fF, respectively.

First, we simulate the writing scheme shown in Figure 2(b) using a 2×22 \times2 FeFET memory array with VW=3.6V_{\mathrm{W}} = 3.6 V, which is divisible by three and ensures that 1/3⋅VW1/3 \cdot V_{\mathrm{W}} is below the polarization threshold voltage VPTHV_{\mathrm{PTH}}, meaning that polarization states of other cells will not be disturbed. The VPR for all FeFETs is set to 4.5. VPR = 4.5 in Equation (5) can be treated as if the speed was not calibrated based on a unit-distance but on a maximum polarization amplitude range. All other parameters remain at their default values as per Table I. Figure 8 illustrates the writing scheme for all memory cells in the array. All FeFETs start in the high-VTH state, and a specific set of voltage pulses is applied to the array at different time steps. The writing operation is performed as described in § 2-B and all necessary signals can be found in Figure 8(a). The length of the pulse determines the polarization level. The first VW pulse is applied to WL0 and targets the FeFET at array coordinates (0,0). The second set of voltage pulses lasts longer than the first one, resulting in a higher polarization state for FeFET at array coordinates (1,0). FeFET at array coordinates (0,1) exhibits the lowest level of polarization since the third set of pulses, which enables its writing, is shorter than both previous ones. The last set of voltage pulses is designed to replicate the first one. Therefore, the polarization state of FeFET at array coordinates (1,1) is the same as for the FeFET at array coordinates (0,0), demonstrating the stability of the polarization logic within the NN-based transistor model.

Input signals and FeFET polarization levels

Figure 8. Demonstration of selective write operations on a FeFET memory array using the proposed NN-based compact model. Memory array inputs (a) switch the polarization state of individual transistors (b). For clarity, a 2×22 \times2 array configuration is depicted; however, the model was validated for array sizes up to 512×512512 \times512.

In contrast to the memory writing procedure, which primarily reflects the polarization logic correctness and behavior, memory reading evaluates the drain current calculation governed by the

NN within the FeFET model. To ensure a moderate level of simulation complexity, we employ a 128×128128 \times128 memory array with parasitics. For visual clarity, we simulate a 3-bit memory array with 8 states. First, we set the operating voltage to 1 V and identify 8 polarization states corresponding to 8 equidistant drain currents in the FeFET with the fixed VDS=VGS=1V_{\mathrm{DS}} = V_{\mathrm{GS}} = 1 V. Then, each FeFET in the 128×128128 \times128 memory array is assigned a random initial state VP0 from the calculated set of polarization states. Figure 9 illustrates the memory read simulation, where some selected input and output signals are shown. The read pulse of VR=1V_{\mathrm{R}} = 1 V is applied to WL at different points in time, enabling the memory read for different rows of the array. Since constant voltage of 1 V and 0 V is applied to BL and SL, respectively, WL pulses in Figure 9(a) activate memory read in the first 8 rows of the array. Outputs in Figure 9(b) correspond to array columns which are selected randomly to cover the entire range of memory outputs. BL currents in Figure 9(b) have 8 clearly distinguishable levels which correspond to the selected states of a 3-bit memory. Although all levels are consistent across multiple bit lines, the transition behavior slightly differs across these lines, which is not visible due to scale. The difference in transition behavior is caused by parasitic elements in the array, which slow down the transition. In particular, bit lines far apart in the array layout will exhibit the most difference in transition behavior due to a large number of accumulated parasitic elements.

Read voltages and current representing stored 3-bit memory values

Figure 9. SPICE simulation for memory read in a 128×128128 \times128 memory array using the NN-based FeFET model. VP0 for each transistor is initialized randomly from 8 states corresponding to a 3-bit memory. Each input (top) receives a read voltage pulse and produces an output in the form of a current (bottom).

Another crucial aspect of the transistor model performance is the simulation time. For simulation time measurements, we employ the memory array architecture from previous experiments and vary its dimensions. All parasitic elements are also included to increase the simulation’s complexity. The general setup for the experiment and input signals is very similar to that of the previous simulation, with the 128×128 array shown in Figure 9. In contrast to the previous experiment, VP0 was not selected from a 3-bit set but was initialized randomly with values in the range [-1, 1]. The transient simulation time is fixed at 15 ns across all simulations to equalize conditions. Figure 10 illustrates how the size of the memory array affects SPICE simulation time with the NN-based FeFET model. Since we employ square M×M memory arrays, the number of transistors increases quadratically, and the simulation time increases exponentially. All simulations converged without any issues with the same simulation options across all array sizes. A 512×512 memory array with 262 144 FeFETs and more than 1.5 million parasitic elements takes around 15 minutes to simulate. Disabling polarization logic in the model via a VPMOD switch reduces simulation time for large memory arrays by up to 17%.

SPICE simulation time for M×M memory arrays with the Verilog-A NN-based FeFET model

Figure 10. SPICE simulation time for M×M memory arrays with the Verilog-A NN-based FeFET model. Polarization recalculation is enabled for blue and disabled for red. Polarization state VP0 is initialized randomly for each transistor.

Conclusion

In this work, we presented a derivative-free neural network NN-based modeling approach aimed to accelerate early-stage circuit prototyping for emerging transistor technologies using only measurement data. By incorporating spectral normalization, our method effectively mitigates complex measurement noise and ensures smooth derivative behavior without relying on numerical solver gradients during training. We validated our approach on a 28 nm FeFET, accurately capturing a continuous range of operating voltage conditions and polarization amplitudes. Furthermore, the model’s scalability and compatibility with standard SPICE engines were demonstrated through transient simulations of large-scale memory arrays with up to 512×512 size and parasitic effects. By bypassing the prolonged, expert-intensive process of traditional transistor model development, our approach provides a fast and reliable path from raw measurement data to circuit simulation, significantly accelerating the deployment cycle for next-generation semiconductor devices.

Acknowledgment

We would like to thank Halid Mulaosmanovic, Stefan Dünkel, and Sven Beyer from GlobalFoundries for their valuable support and for providing the FeFET wafers for measurements.

References

  1. [1]Y. S. Chauhan, S. Venugopalan, M.-A. Chialkiadaki, M. A. U. Karim, H. Agarwal, S. Khandelwal, N. Paydavosi, J. P. Duarte, C. C. Enz, A. M. Niknejad, and C. Hu, “Bsim6: Analog and rf compact model for bulk mosfet,” IEEE Transactions on Electron Devices, vol. 61, no. 2, pp. 234–244, 2014.DOI
  2. [2]C.-T. Tung and C. Hu, “Neural network-based bsim transistor model framework: Currents, charges, variability, and circuit simulation,” IEEE Transactions on Electron Devices, vol. 70, no. 4, pp. 2157–2160, 2023.DOI
  3. [3]J. Wang, Y.-H. Kim, J. Ryu, C. Jeong, W. Choi, and D. Kim, “Artificial neural network-based compact modeling methodology for advanced transistors,” IEEE Transactions on Electron Devices, vol. 68, no. 3, pp. 1318–1325, 2021.DOI
  4. [4]S. Eom, H. Yun, H. Jang, K. Cho, S. Lee, J. Jeong, and R. Baek, “Neural compact modeling framework for flexible model parameter selection with high accuracy and fast spice simulation,” Advanced Intelligent Systems, vol. 6, 01 2024.DOI
  5. [5]R. Novkin and H. Amrouch, “Dual neural network framework with spice integration for fast and accurate transistor modeling,” Advanced Intelligent Systems, vol. 7, 04 2025.DOI
  6. [6]T. S. Böschke, J. Müller, D. Bräuhaus, U. Schröder, and U. Böttger, “Ferroelectricity in hafnium oxide thin films,” Applied Physics Letters, vol. 99, no. 10, p. 102903, 09 2011. [Online]. Available: https://doi.org/10.1063/1.3634052
  7. [7]——, “Ferroelectricity in hafnium oxide: Cmos compatible ferroelectric field effect transistors,” in 2011 International Electron Devices Meeting, 2011, pp. 24.5.1–24.5.4.
  8. [8]M. Jerry, P.-Y. Chen, J. Zhang, P. Sharma, K. Ni, S. Yu, and S. Datta, “Ferroelectric fet analog synapse for acceleration of deep neural network training,” in 2017 IEEE International Electron Devices Meeting (IEDM), 2017, pp. 6.2.1–6.2.4.DOI
  9. [9]D. Reis, K. Ni, W. Chakraborty, X. Yin, M. Trentzsch, S. D. Dünkel, T. Melde, J. Müller, S. Beyer, S. Datta, M. T. Niemier, and X. S. Hu, “Design and analysis of an ultra-dense, low-leakage, and fast fet-based random access memory array,” IEEE Journal on Exploratory Solid-State Computational Devices and Circuits, vol. 5, no. 2, pp. 103–112, 2019.DOI
  10. [10]A. Mema, S. Thomann, and H. Amrouch, “Fefet-based cmos current starvation programmable delay element,” IEEE Solid-State Circuits Letters, vol. 9, pp. 166–169, 2026.DOI
  11. [11]A. Mema, S. Thomann, N. S. Dhakad, and H. Amrouch, “First demonstration of 28 nm fabricated fefet-based nonvolatile 6t sram,” IEEE Transactions on Circuits and Systems I: Regular Papers, 2026.DOI
  12. [12]S. Thomann, R. Novkin, J. Li, Y. Hu, J. Xiong, and H. Amrouch, “Ml-tcad: Accelerating fefet reliability analysis using machine learning,” IEEE Transactions on Electron Devices, vol. 71, no. 1, pp. 213–222, 2024.DOI
  13. [13]Y.-M. Tseng, J.-T. Wu, W.-T. Chen, C.-H. Wu, and W. P.-H. Hu, “Physics-aware graph neural network for fefet variability analysis: Enabling 10^7× speedup, r2 > 0.98 accuracy, and efficient transfer learning to new architectures,” in 2025 IEEE International Electron Devices Meeting (IEDM), 2025, pp. 1–4.DOI
  14. [14]T. Miyato, T. Kataoka, M. Koyama, and Y. Yoshida, “Spectral normalization for generative adversarial networks,” ArXiv, vol. abs/1802.05957, 2018.arxiv.org/abs/1802.05957
  15. [15]H. Mulaosmanovic, S. Dünkel, M. Trentzsch, S. Beyer, E. T. Breyer, T. Mikolajick, and S. Slesazeck, “Investigation of accumulative switching in ferroelectric fets: Enabling universal modeling of the switching behavior,” IEEE Transactions on Electron Devices, vol. 67, no. 12, pp. 5804–5809, 2020.DOI
  16. [16]J. Park, H. Kim, H. Cho, and J. Jeon, “Nsip: Neural network spice integration platform for ferroelectric field-effect transistor-based crossbar arrays,” Advanced Intelligent Systems, vol. 7, 06 2025.DOI

Paper details

vlnc:W_a9bgttqc
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents